JP4894747B2 - Partial region detection device, object identification device, and program - Google Patents

Description

  The present invention relates to a partial region detection device, an object identification device, and a program, and more particularly, to a partial region detection device that detects a partial region for identifying an identification target, and an image in which an identification target image captures an identification target. It is related with the target object identification apparatus which determines whether it is.

  2. Description of the Related Art Conventionally, an object type determination device that identifies an area where an object exists and determines the type of an object using an edge direction or intensity distribution in the area as a feature amount is known (Patent Document 1).

In addition, a method is known in which a small region is set by dividing an input pattern image into a lattice shape, a histogram in the edge direction is calculated for each small region, and a person is identified by using the histogram as a feature (Non-Patent Document 1). ).
JP 2007-156626 A "Histogram of orientated gradients for human detection", IEEE CVPR2005

  However, in the technique described in Patent Document 1 described above, the edge direction and the intensity distribution as the feature amount are calculated for the target region, and the position information in the target region is not reflected in the feature amount. There is a problem that.

  Further, in the technique described in Non-Patent Document 1, since the method of dividing the small area is regular, if the target object partially changes, the feature amount changes, and the identification performance deteriorates. There is a problem that.

  The present invention has been made in order to solve the above-described problems, and a partial region detection apparatus and a target capable of accurately identifying whether an identification target image is an image obtained by capturing an identification target object. It is an object to provide an object identification device and a program.

  In order to achieve the above object, the partial region detection device according to the first aspect of the present invention relates to a case where each pixel of the image is a pixel of interest for each of a plurality of images obtained by imaging the identification target object. A feature amount calculating unit that calculates a feature amount that is obtained based on a density difference between peripheral pixels that exist around the target pixel and that indicates a degree of change in density in a predetermined region that includes the target pixel and the peripheral pixel; Average total calculation means for calculating an average value of the feature quantities or a sum of the feature quantities for each target pixel based on the feature quantities of each of the plurality of images calculated by the feature quantity calculation means; Partial area detection means for detecting, as a partial area for identifying the identification target object, a partial area including pixels where the average value or the total sum calculated by the average total calculation means is maximum. In is configured.

  According to a second aspect of the present invention, there is provided a program that includes, for each of a plurality of images obtained by capturing an image of an identification target by a computer, a periphery existing around a target pixel when each pixel of the image is a target pixel A feature amount calculation unit that calculates a feature amount that is obtained based on a density difference between pixels and that indicates a degree of change in density in a predetermined region including the target pixel and the surrounding pixels, and is calculated by the feature amount calculation unit Based on the feature amount of each of the plurality of images, the average total calculation unit that calculates the average value of the feature amount or the sum of the feature amount for each target pixel, and the average total calculation unit For functioning as a partial area detecting means for detecting a partial area including pixels where the average value or the sum is maximized as a partial area for identifying the identification object Is a program.

  According to the first invention and the second invention, for each of a plurality of images obtained by imaging the identification object by the feature amount calculation unit, the pixel of interest when each pixel of the image is the pixel of interest A feature amount that is obtained based on a density difference between peripheral pixels existing around the pixel and that indicates a degree of change in density in a predetermined area including the target pixel and the peripheral pixels is calculated. Based on the feature amounts of each of the plurality of images calculated by the feature amount calculation unit, the average sum calculation unit calculates the average value of feature amounts or the sum of the feature amounts for each target pixel.

  Then, the partial region detection unit detects a partial region including pixels whose average value or total sum calculated by the average total calculation unit is a maximum as a partial region for identifying the identification target object.

  In this way, the identification object to be imaged is detected by detecting the partial area including the pixel whose average value or total sum of the feature amount indicating the degree of change in density is the maximum as the partial area for identifying the identification object. It is possible to detect a partial area that can reduce the influence of the change in the feature amount of the image, and use the detected partial area to increase whether the identification target image is an image obtained by imaging the identification target object. Can be identified with accuracy.

  The partial area detecting means according to the first invention can detect a partial area centering on a pixel having a maximum average value or total sum as a partial area for identification. As a result, it is possible to detect a partial region that can further reduce the influence of the change in the imaged identification object on the feature amount of the image.

  The average total calculation means according to the first invention calculates an average value of feature quantities or a sum of feature quantities for each target pixel based on the feature quantities of each of the plurality of images calculated by the feature quantity calculation means. The feature amount image having the average value or the sum total as the pixel value is generated, and the partial region detection unit uses the pixel having the maximum pixel value of the feature amount image or a pixel around the pixel having the maximum value as the target pixel. The feature amount is calculated, and the partial region including the pixel having the maximum pixel value of the feature amount image, and the smaller partial region as the calculated feature amount is larger, is used for identifying the identification target object. It can be detected as a partial region. As a result, it is possible to detect a partial region having a size corresponding to the magnitude of the change in the identification object.

  The partial area detecting means can detect a plurality of partial areas for identification. Thereby, when there are a plurality of pixels having a maximum pixel value, a plurality of partial regions can be detected.

  The feature amount may be a gradient magnitude. Depending on the magnitude of the gradient, it is possible to indicate the degree of change in density in a predetermined region including the target pixel and the peripheral pixels.

  The identification object may be an object whose position or size varies. As a result, even if the position or size of the part of the identification target object varies, it is possible to detect a partial region in which the influence of the variation of the identification object on the image feature amount can be reduced.

  The identification object can be an object whose position or size differs depending on the type. As a result, even if the position or size of the part of the identification target object differs depending on the type, it is possible to detect a partial region that can reduce the influence of the difference in the part of the identification target object on the feature amount of the image.

  An object identification device according to a third aspect of the present invention is stored in the storage unit with respect to a storage unit that stores the position and size of the partial region detected by the partial region detection device, and an identification target image. A partial area setting means for setting a partial area having the same position and size as the position and size of the partial area; and the feature amount of the partial area of the identification target image set by the partial area setting means A feature amount calculation means for calculating the feature amount, the feature amount of the detected partial area of the image obtained by imaging the identification object obtained in advance, and the feature amount calculation means calculated by the feature amount calculation means A determination unit configured to compare the feature quantity of the partial region and determine whether the identification target image is an image of the identification target object based on a comparison result. To have.

  A program according to a fourth invention is stored in the storage means for a storage means that stores the position and size of the partial area detected by the program according to the second invention, and an image to be identified. The partial area setting means for setting a partial area having the same position and size as the position and size of the partial area, and the feature amount of the partial area of the identification target image set by the partial area setting means. The feature amount calculation means to calculate, the feature amount of the detected partial area of the image obtained by imaging the identification object obtained in advance, and the portion calculated by the feature amount calculation means A determination unit that compares the feature quantity of the region and determines whether the identification target image is an image of the identification target object based on a comparison result; Is a program for making the function Te.

  According to the third invention and the fourth invention, the position and size of the partial area for identifying the detected identification object is stored in the storage means, and the partial area setting means Then, a partial area having the same position and size as that of the partial area stored in the storage means is set.

  Then, the feature amount calculation unit calculates the feature amount of the partial region of the identification target image set by the partial region setting unit. Compare the feature quantity of the detected partial area of the image obtained by imaging the identification object obtained in advance by the determination means and the feature quantity of the partial area calculated by the feature quantity calculation means, and compare Based on the result, it is determined whether the identification target image is an image obtained by imaging the identification target object.

  In this way, for the identification target image, a partial area that can reduce the influence of a change in the captured identification target on the feature amount of the image is set, and using the feature amount of the partial area, the identification target image is It is possible to identify with high accuracy whether the image is an image of the identification object.

  As described above, according to the partial region detection apparatus and the program of the present invention, in order to identify the identification target object, the partial region including the pixel whose average value or total sum of the feature amounts indicating the degree of change in density is maximized is identified. By detecting this as a partial area, it is possible to detect a partial area that can reduce the effect of a change in the imaged identification object on the feature quantity of the image. It is possible to obtain an effect that it is possible to identify with high accuracy whether or not the image is an image of the identification target.

  According to the object identification device and the program of the present invention, a partial area that can reduce the influence of the change of the identification object to be captured on the feature quantity of the image is set for the identification target image, and the feature quantity of the partial area is set. Using this, it is possible to identify with high accuracy whether or not the identification target image is an image obtained by imaging the identification target.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In addition, the case where this invention is applied to the target object identification apparatus which identifies whether a captured image is an image which imaged the person as an identification target object is demonstrated to an example.

  As shown in FIG. 1, an object identification device 10 according to the present embodiment is attached to a vehicle (not shown), and captures an image of the front of the vehicle to generate an image, and the imaging device 12. The computer 14 for identifying whether or not the captured image obtained from the image is a captured image of a person, and the display device 16 for displaying the identification result of the computer 14 are provided. The object identification device 10 corresponds to the partial region detection device and the object identification device of the present invention.

  The imaging device 12 images the front of the vehicle and generates an image signal of the image (not shown), and an A / D conversion unit (not shown) that A / D converts the image signal generated by the imaging unit. And an image memory (not shown) for temporarily storing the A / D converted image signal.

  The computer 14 includes a CPU, a RAM, and a ROM that stores a program for executing each of a learning processing routine and an object identification processing routine described later, and is functionally configured as follows. . The computer 14 stores a learning image database 20 that stores a plurality of learning images obtained by imaging a person, a gradient calculation unit 22 that calculates the gradient magnitude for each pixel of the learning image, and the gradient magnitude of each pixel of the learning image. A gradient image generation unit 24 that generates a gradient image using the average value of the gradient magnitude as a pixel value on an average basis for each pixel, and detects each point at which the pixel value becomes a maximum value in the gradient image, and the maximum point Each of which is a central position, and a partial region determination unit 26 that determines a partial region having a size corresponding to the gradient of the maximum point in the gradient image, and the position of each partial region determined by the partial region determination unit 26 And a partial region information storage unit 28 for storing information indicating the size and a histogram of each gradient direction (hereinafter referred to as gradient gradient) from each determined partial region in the learning image. A feature vector indicating a gradient histogram is generated for each partial region, and data obtained by averaging the feature vectors of each partial region for a plurality of learning images is calculated as a reference feature amount. A calculation unit 30 and a reference feature amount storage unit 32 that stores a feature vector indicating a gradient histogram that is a calculated reference feature amount for each partial region are provided.

  The learning image database 20 stores n images {x1,..., Xn} of various sizes obtained by cutting out an area representing a person from an image obtained by capturing a person as learning images. The learning images all have the same aspect ratio, and the aspect ratio according to the aspect ratio of the person.

  The gradient calculation unit 22 calculates the gradient size of each pixel for each of the plurality of learning images as described below. First, the learning image acquired from the learning image database 20 is normalized to a predetermined size to generate a normalized image {x1 bar,..., Xn bar}. Next, the magnitude of the gradient is calculated for each pixel of each normalized image according to the following equation (1).

  However, m (u, v) represents the magnitude of the gradient at the image coordinates (u, v), and the bar (u, v) of x represents the image coordinates (u, v) of the bar of the normalized image x. Represents a pixel value.

  Here, as shown in the above equation (1), the magnitude of the gradient is obtained based on a density difference between peripheral pixels existing around the target pixel when each pixel of the image is the target pixel. And the change degree of the density | concentration in the predetermined area | region containing a surrounding pixel is shown.

  As shown in FIG. 2A, the gradient image generation unit 24 generates a gradient image {g1,..., Gn} having the gradient magnitude calculated by the gradient calculation unit 22 as a pixel value. The average of the pixel values for each pixel of the generated n gradient images is calculated according to the equation (2), and the average value of the pixel values of the gradient image as shown in FIG. A gradient average image G to be generated is generated.

  Here, G (u, v) represents a pixel value at the image coordinates (u, v) of the gradient average image G, and gk (u, v) represents the image coordinates (u, v) of the kth gradient image gk. ) Represents the pixel value.

  The partial area determination unit 26 determines a partial area for identifying whether or not the image is an image of a person, as will be described below, from the generated gradient average image G. First, the magnitude of the gradient of each pixel of the gradient average image G is calculated, and a gradient image Gg having the pixel magnitude as the gradient magnitude of each pixel calculated from the gradient average image G is generated. Note that the gradient size calculation method may be the same as that used when generating the gradient image.

  Then, as shown in FIG. 3A, each pixel having a maximum pixel value in the gradient average image G is extracted. Whether or not the pixel value is a maximum value is determined by whether or not the pixel value of the target pixel is the maximum within a predetermined window (for example, a 5 × 5 window) centered on the target pixel. Finally, for each extracted maximum point, the position and size of the partial region for identification are determined based on the position of the maximum point of the pixel value and the pixel value of the gradient image Gg of the gradient average image G. For example, as shown in FIG. 3B, the position of the partial region is centered on the position of the maximum point, and the size of the partial region is the pixel of the pixel of the gradient image Gg corresponding to the pixel near the maximum point. The larger the value is, the smaller the size of the partial region is determined, and the smaller the pixel value of the pixel of the gradient image Gg corresponding to the pixel near the maximum point is, the larger the size of the partial region is determined (for example, FIG. In the example of B), since the pixel value of the pixel of the gradient image Gg is smaller in the partial region 1, the size of the partial region 1 is determined to be larger, and the pixel of the pixel in the gradient image Gg is determined in the partial region 2. Since the value is large, the size of the partial region 2 is determined to be small.)

  As described above, the position and size of the partial region are determined for each of the maximum points of the pixel values of the gradient average image G, and information indicating the position and size of each partial region is stored in the partial region information storage unit 28. Remembered.

  The reference feature amount calculation unit 20 normalizes the learning image to a specified size, sets the partial region having the position and size determined as described above, and normalizes the normalized learning image. As shown in FIG. 4, the gradient magnitude and gradient direction of each pixel are calculated for the set partial region in the learned image. The magnitude of the gradient is calculated according to the above equation (1), and the direction of the gradient is calculated according to the following equation (3).

  Where θ (u, v) represents the direction of the gradient in pixel coordinates (u, v), and x bar (u, v) represents the image coordinates (u, v) of the normalized learning image x bar. ) Represents the pixel value.

  Then, based on the calculated gradient magnitude and gradient direction of each pixel, a gradient histogram for the partial region is calculated as shown in FIG. At the time of voting the histogram, the gradient histogram is calculated by voting with the magnitude of the gradient.

  As described above, a gradient histogram for each partial region is calculated for each normalized learning image, a feature vector indicating the gradient histogram is generated for each partial region, and a plurality of feature vectors for each partial region are learned. Data averaged for the image is calculated as a reference feature value, and the reference feature value is stored in the reference feature value storage unit 32 for each partial region.

  Next, the principle of this embodiment will be described. Conventionally, when an identification object is detected from an image pattern using learning type identification, in order to cope with a minute positional shift while retaining shape information of the entire image, each small region obtained by dividing the image into a lattice shape is used. Then, feature vectors such as edge direction histograms are calculated to generate feature vectors. However, when detecting an object that is partially deformed like a person and the position or size of each part varies, the position change of each part of the object compared to the size of the divided small area If is large, the feature vectors are greatly different, and it is difficult to detect the object. Therefore, in the present embodiment, the maximum point of the pixel value of the gradient average image is obtained from the learning image as the fluctuation center position of each part, and the maximum point of the pixel value is used as an index indicating the magnitude of the fluctuation of each part. The magnitude of the gradient in the surrounding gradient average image is obtained, and the position and size of the partial region are determined. Thereby, it is possible to reduce the influence of the partial position fluctuation of the object on the feature amount, and it is possible to realize robust object detection.

  In addition, the computer 14 of the object identification device 10 acquires an image acquired from the imaging device 12 and the captured image acquired by the image acquisition unit 40 at each position of the captured image. Information indicating the size and position of each partial area stored in the partial area information storage unit 28, and an identification target image extraction unit 42 that sets an identification target area and extracts an image of each identification target area as an identification target image On the basis of the extracted identification target image, a partial region setting unit 44 that sets each partial region determined by learning, and a feature that indicates a gradient histogram by calculating a gradient histogram for each of the set partial regions A gradient histogram calculation unit 46 for generating a vector, a feature vector indicating a gradient histogram calculated for a partial region of the identification target image, and a reference A feature vector indicating a gradient histogram as a reference feature amount of the same partial area stored in the collection amount storage unit 32 is compared to determine whether the identification target image is an image of a person, and the identification target And an identification unit 48 for identifying an image.

  The identification target image extraction unit 42 sets an identification target region at various positions in the captured image in various sizes with the same aspect ratio as the learning image, and sets the image of the identification target region to the captured image. Extracted as an identification target image. For example, as shown in FIGS. 6A and 6B, for one captured image, areas of different sizes are set at each position of the image, and an identification target image is extracted from each set area. To do.

  The partial region setting unit 44 changes the identification target image extracted from the captured image to a specified size, normalizes the size, and stores the normalized identification target image in the partial region information storage unit 28. Each partial area is set based on the position and size of each partial area determined by the stored learning. In this way, by setting a partial area for the normalized identification target image, a partial area having the same position and size as each of the positions and sizes of the plurality of partial areas determined by learning is obtained. , It can be set for the identification target image.

  The gradient histogram calculation unit 46 calculates and calculates the gradient magnitude and gradient direction of each pixel for the set partial region in the normalized identification target image in the same manner as the reference feature amount calculation unit 30 described above. A gradient histogram of the partial region is calculated based on the gradient magnitude and gradient direction of each pixel. Further, a feature vector indicating the calculated gradient histogram is generated. A gradient histogram is calculated for each partial region, and a feature vector indicating the gradient histogram is generated for each partial region.

  The identification unit 48 includes a feature vector indicating a gradient histogram as a reference feature amount obtained by learning stored in the reference feature amount storage unit 32, and a feature vector indicating a gradient histogram calculated by the gradient histogram calculation unit 46. The feature vector indicating the calculated gradient histogram is compared with the reference feature amount for each partial region, and based on the comparison determination result of each partial region, the identification target image is a person. It is determined whether the image is a captured image, and an identification target image is identified.

  Next, the operation of the object identification device 10 according to the present embodiment will be described. First, for each of a plurality of images obtained by imaging a person, an area representing the person is cut out with an aspect ratio determined according to the aspect ratio of the person, and n pieces of various sizes are cut out. Are stored in the learning image database 20 as learning images.

  Then, a learning process routine shown in FIG. 7 is executed in the computer 14. First, in step 100, a plurality of learning images representing a person are acquired from the learning image database 20, and in step 102, the size of each of the learning images acquired in step 100 is normalized, and a plurality of normal images are obtained. Generate a digitized image.

  In step 104, for each normalized image generated in step 102, the gradient magnitude of each pixel is calculated according to the above equation (1), and the gradient magnitude of each pixel is used as the pixel value. Generate multiple images. In the next step 106, the pixel values of the respective pixels are averaged for the plurality of gradient images generated in step 104 to generate a gradient average image. In step 108, the gradient average image generated in step 106 is generated. For each of the pixels, the magnitude of the gradient is calculated according to the above equation (1), and a gradient image with the calculated gradient magnitude as the pixel value is generated.

  In step 110, one or more local maximum points at which the pixel value is maximum are extracted from the gradient average image generated in step 106. In step 112, each local maximum point extracted in step 110 is extracted. The position and size of the partial region are determined based on the position of the maximum point and the size of the pixel value of the pixel corresponding to the extracted maximum point in the gradient image of the gradient average pixel generated in step 108 above. To do. In step 114, information indicating the position and size of each partial area determined in step 112 is stored in the partial area information storage unit 28.

  In step 116, the partial regions determined in step 112 are set for each of the plurality of normalized images generated in step 102. In step 118, the above steps are performed for each normalized image. A gradient histogram for each partial region set in 116 is calculated, and a feature vector indicating the calculated gradient histogram for each partial region is generated. In the next step 120, the feature vector indicating the gradient histogram for each partial region calculated in step 118 is averaged for a plurality of normalized images, and the gradient histogram is used as a reference feature amount for each partial region. Average data of the indicated feature vectors is calculated, the reference feature value is stored for each partial region in the reference feature value storage unit 32, and the learning processing routine is terminated.

  Then, when a predetermined area in front of the vehicle is imaged by the imaging device 12 while the vehicle equipped with the object identification device 10 is traveling, the object identification processing routine shown in FIG.

  First, in step 130, a captured image is acquired from the imaging device 12, and in step 132, as shown in FIGS. 6A and 6B, each size is determined from each position of the captured image acquired in step 130. Are extracted as identification target images. In step 134, each of the identification target images extracted in step 132 is normalized to a prescribed size, and in the next step 136, the position and size of each partial area are read from the partial area information storage unit 28. Get information about.

  In step 138, a variable n for identifying the identification target image is set to an initial value of 1. In step 140, the identification target image n normalized in step 134 is compared with the identification target image n in step 136. Each partial area is set based on the acquired information on each partial area.

  In step 142, the variable s for identifying the partial area is set to 1 which is an initial value, and in the next step 144, the partial area s set in step 140 of the normalized identification target image n. For each of the pixels, the magnitude and direction of the gradient are calculated according to the above equations (1) and (3).

  In step 146, a gradient histogram is calculated for the partial region s based on the magnitude and direction of the gradient calculated in step 144, and a feature vector indicating the calculated gradient histogram is generated. In step 148, The feature vector indicating the gradient histogram as the reference feature amount of the partial region s is compared with the feature vector indicating the gradient histogram calculated in step 146, and the feature vector indicating the calculated gradient histogram is the partial region s. It is determined whether or not it is similar to the reference feature amount.

  In the next step 150, it is determined whether or not the variable s is less than a constant S indicating the number of partial regions, and the partial region in which the variable s is less than the constant S and the gradient histogram comparison determination is not performed. If there is, the variable s is incremented in step 152 and the process returns to step 144. On the other hand, when the variable s is equal to or greater than the constant S in step 150 and the gradient histogram comparison determination is performed for all partial regions, in step 154, all the comparison determination results in step 148 are displayed. Based on this, it is determined whether or not the identification target image n is an image of a person. In step 154, for example, when it is determined that the feature vector is similar to the reference feature amount based on a comparison determination result of a predetermined ratio or more, the identification target image n is identified as an image obtained by capturing a person.

  In step 156, it is determined whether or not the identification target image n is identified as an image in which a person is imaged in step 154. If it is identified as an image in which a person is imaged, step 158 is performed. Then, the identification result indicating that the captured image is an image of a person is displayed on the display device 16, and the object identification processing routine is terminated.

  On the other hand, if it is determined in step 156 that the image is not an image of a person, whether or not the variable n is less than a constant N indicating the number of identification target images extracted in step 132 is determined in step 160. If the variable n is less than the constant N and there is an identification target image that has not been subjected to the above-described identification processing, the variable n is incremented in step 162 and the process returns to step 140. On the other hand, if the variable n is equal to or greater than the constant N in step 160 and the above identification processing is performed for all the identification target images, the captured image is not an image of a person captured in step 164. The identification result is displayed on the display device 16, and the object identification processing routine is terminated.

  As described above, according to the object identification device according to the present embodiment, a partial region centered on a pixel having an average gradient magnitude maximum in a plurality of learning images obtained by imaging a person is represented as a person. Even if the position or size of the part of the person to be imaged fluctuates, the influence of the fluctuation of the person on the feature amount of the image is detected. Can be detected.

  In addition, as an index indicating the magnitude of the variation of the person, the gradient of the maximum point in the gradient average image is obtained, and the partial region having a size corresponding to the gradient of the maximum point is detected. It is possible to detect a partial area having a size corresponding to the magnitude of the variation of the person within.

  In addition, a partial area that can reduce the influence of fluctuation of a person to be captured on the feature quantity of the image is set with respect to the identification target image, and the identification target image is identified by using the feature quantity vector of the partial area. It is possible to identify with high accuracy whether or not the image is an image of a person. In addition, it is possible to robustly detect a person as an identification object whose position or size of a part varies.

  In the above embodiment, the object identification device detects a partial region for identifying a person by learning and identifies whether the identification target image is an image of a person. Although described as an example, the present invention is not limited to this, and a partial area detection device that detects a partial area for identifying a person through learning and whether or not the identification target image is an image of a person is identified. You may comprise separately the target object identification apparatus to perform. In this case, information on the partial region and the reference feature amount may be obtained by the partial region detection device, and the information on the partial region and the reference feature amount may be stored in the memory of the object identification device.

  Moreover, although the case where the magnitude of the gradient is averaged for a plurality of learning images and the maximum point is extracted has been described as an example, the present invention is not limited to this. A maximum point may be extracted. In this case, for a plurality of learning images, the sum of gradient magnitudes is calculated for each pixel, and a gradient sum image having the calculated sum of gradient magnitudes as pixel values is generated. What is necessary is just to extract the point where becomes the maximum.

  Further, although the case where the identification target is a person has been described as an example, the present invention is not limited to this, and an object other than a person whose position or size varies may be used as the identification target. Further, not only an object such as a person that partially fluctuates, but also an object whose position and size are different depending on the type may be used as the identification target object. For example, when the vehicle is an identification object, as shown in FIGS. 9A and 9B, even if the size or position of a part such as a lamp is different depending on the vehicle type, the partial region is changed depending on how the part is different. By setting, the identification object can be robustly identified.

  Further, the case where the partial area is set for the identification target image whose size is normalized has been described as an example, but the present invention is not limited to this, and the position of each of the partial areas determined by learning and A partial area having the same position and size as the size may be set in the identification target image.

  Further, the case where the gradient magnitude calculated by the above equation (1) is used as the feature amount indicating the degree of density change has been described as an example. However, the present invention is not limited to this. A feature value obtained based on a density difference between peripheral pixels existing around the target pixel when the target pixel is used, and indicating a degree of change in density in a predetermined area including the target pixel and the peripheral pixel May be used.

  In addition, a feature vector is obtained as a reference feature amount for each partial area from the learning image, the feature vector of each partial area of the identification target image is compared with the reference feature amount of each partial area, and the identification target image captures a person The case of identifying whether or not the image is an example has been described as an example. However, the present invention is not limited to this, and the distribution of feature vectors obtained from each of a learning image that captures a person and a learning image that does not capture a person And a feature vector obtained from the identification target image may be compared to identify whether the identification target image is an image of a person. In this case, after determining a partial area from a learning image (person learning image) obtained by imaging a person, a gradient histogram is calculated in each partial area of the learning image. Then, for each person learning image, a feature vector is generated using all the calculated gradient histograms of each partial region (for example, a feature vector in which values of gradient histograms of all partial regions are arranged in order) is generated. A learning image (non-person learning image) obtained by imaging a non-person (non-person) is also prepared, and a gradient histogram of each partial region is similarly calculated for the non-person learning image. Then, for each non-human learning image, a feature vector is generated using all the calculated gradient histograms of the partial regions. Then, the feature vector distribution is learned and stored from the feature vectors of all the person learning images, and the feature vector distribution is learned and stored from the feature vectors of all the non-person learning images. During the identification process, it is determined whether the feature vector calculated in the same manner for the identification target image belongs to the distribution of the feature vector of the person learning image or the distribution of the feature vector of the non-person learning image. Identify whether or not a person is represented. Note that nearest neighbor identification, linear discrimination, a support vector machine, or the like may be used as a learning or identification method for the distribution of feature vectors.

It is the schematic which shows the structure of the target object identification apparatus which concerns on embodiment of this invention. (A) It is an image figure which shows the magnitude | size of the gradient calculated about the learning image, (B) It is an image figure which shows a gradient average image. (A) The image figure which shows the maximum point extracted from the gradient average image, (B) The image figure which shows the determined partial area | region. It is an image figure which shows a mode that the magnitude | size and position of the gradient were calculated about each pixel in a partial area. It is a graph which shows a gradient histogram. It is an image figure which shows a mode that an identification object image is extracted from each position and each magnitude | size from a captured image. It is a flowchart which shows the content of the learning process routine in the computer of the target object identification apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the target object identification process routine in the computer of the target object identification device which concerns on embodiment of this invention. It is an image figure which shows a mode that the magnitude | size and position of parts, such as a lamp, differ with vehicle models about the vehicle as an identification target object.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Object identification apparatus 12 Imaging device 14 Computer 20 Learning image database 20 Reference | standard feature-value calculation part 22 Gradient calculation part 24 Gradient image generation part 26 Partial area | region determination part 28 Partial area information storage part 30 Reference | standard feature-value calculation part 32 Reference | standard feature-value Storage unit 42 Identification target image extraction unit 44 Partial region setting unit 46 Gradient histogram calculation unit 48 Identification unit

Claims (12)

For each of a plurality of images obtained by imaging the identification object, is determined based on the density difference of the peripheral pixels existing around the pixel of interest when each pixel of the image is the pixel of interest, and Feature amount calculation means for calculating a feature amount indicating a degree of change in density in a predetermined region including the target pixel and the peripheral pixels;
Average total calculation means for calculating an average value of the feature quantities or a sum of the feature quantities for each target pixel based on the feature quantities of each of the plurality of images calculated by the feature quantity calculation means;
A partial area detecting means for detecting a partial area including pixels in which the average value or the total sum calculated by the average total calculating means is a maximum, as a partial area for identifying the identification target;
A partial area detecting device including:   The partial area detection device according to claim 1, wherein the partial area detection unit detects a partial area centered on a pixel having the maximum average value or the total sum as the partial area for identification. The average total calculation unit calculates an average value of the feature amounts or a sum of the feature amounts for each target pixel based on the feature amounts of each of the plurality of images calculated by the feature amount calculation unit. And generating a feature image having the average value or the sum as a pixel value,
The partial region detection unit calculates the feature amount when a pixel having a maximum pixel value of the feature amount image or a pixel around the pixel having the maximum value is a target pixel, and the pixel value of the feature amount image Or a partial region including a pixel having a maximum value, and detecting a partial region that is smaller as the calculated feature amount is larger as a partial region for identifying the identification target object. 3. The partial area detection device according to 2.   The partial area detection device according to claim 1, wherein the partial area detection unit detects a plurality of partial areas for identification.   The partial region detection apparatus according to claim 1, wherein the feature amount is a gradient.   The partial region detection apparatus according to claim 1, wherein the identification target is an object whose position or size varies.   The partial region detection device according to claim 1, wherein the identification target is an object whose position or size differs depending on the type. Storage means for storing the position and size of the partial area detected by the partial area detection device according to any one of claims 1 to 7,
A partial area setting means for setting a partial area having the same position and size as the position and size of the partial area stored in the storage means for the identification target image;
Feature quantity calculating means for calculating the feature quantity of the partial area of the identification target image set by the partial area setting means;
The feature quantity of the detected partial area of the image obtained by imaging the identification object obtained in advance is compared with the feature quantity of the partial area calculated by the feature quantity calculation means. And determining means for determining whether the identification target image is an image of the identification target based on the comparison result;
An object identification device including: The feature amount calculating means calculates a histogram of the feature amount relating to a density change in each of a plurality of directions in the partial region of the identification target image;
The determination unit includes a histogram of the feature amount relating to a density change in each of the plurality of directions in the detected partial region obtained in advance, and the feature amount in the partial region calculated by the feature amount calculation unit. The object identification device according to claim 8, wherein the object identification device determines whether or not the identification object image is an image of the identification object based on a comparison result. The storage means stores the position and size of each of the detected plurality of partial areas,
The partial area setting means sets, for the identification target image, partial areas that have the same position and size as the positions and sizes of the plurality of partial areas stored in the storage means,
The feature amount calculating unit calculates the feature amount of each of the plurality of partial regions of the identification target image set by the partial region setting unit;
The determination means, for each of the plurality of partial areas, the feature quantity of the detected partial area obtained in advance and the feature quantity of the partial area calculated by the feature quantity calculation means. 10. The object identification device according to claim 8, wherein comparison is made and whether or not the identification target image is an image of the identification target is determined based on a comparison result for each of the plurality of partial regions. Computer
For each of a plurality of images obtained by imaging the identification object, is determined based on the density difference of the peripheral pixels existing around the pixel of interest when each pixel of the image is the pixel of interest, and A feature amount calculating means for calculating a feature amount indicating a degree of change in density in a predetermined region including the target pixel and the peripheral pixels;
Average total calculation means for calculating an average value of the feature quantities or a sum of the feature quantities for each target pixel based on the feature quantities of each of the plurality of images calculated by the feature quantity calculation means; and A program for functioning as a partial area detecting means for detecting a partial area including pixels for which the average value or the total sum calculated by the average total calculating means is a maximum as a partial area for identifying the identification object . Computer
Storage means for storing the position and size of the partial area detected by the program according to claim 11;
A partial area setting means for setting a partial area having the same position and size as the position and size of the partial area stored in the storage means for the identification target image;
The feature amount calculating means for calculating the feature amount of the partial area of the identification target image set by the partial area setting means, and the detection of an image obtained by imaging the identification target object obtained in advance. The feature quantity of the partial area is compared with the feature quantity of the partial area calculated by the feature quantity calculation means, and the identification target image captures the identification target object based on the comparison result A program for functioning as a determination means for determining whether or not an image.

Images